Auto-calibration machine learning algorithms for accurate spike detection from calcium imaging data

Abstract

Deep understanding of action potentials is fundamental to neuroscience research, as it enhances our knowledge of neural dynamics in relation to behavioural and neurological disorders, which is crucial for comprehending the brain and its functions. Calcium imaging is currently an important technique to measure the activity of neurons by monitoring changes of calcium concentration by fluorescent calcium indicators. However, inference of action potentials (spikes) from neuronal calcium imaging data faces several limitations, including the quality of the raw data, the noise level, and the non-linear response of fluorescent indicators towards an increasing number of spikes. Therefore, accurate spike inference requires precise computational calibration models. Auto-calibration adapts to dataset-specific characteristics, eliminating the need for manual tuning of parameters. In this study, we make a quantitative description of spike-evoked calcium transients and employ machine learning algorithms to extract these transients from raw data. In addition, we introduce an auto-calibration method based on hybrid Gaussian mixture models (hybrid GMMs), designed to accurately detect single spikes in datasets recorded using the recently published GCaMP8s/m calcium indicators (genetically encoded calcium sensors with enhanced sensitivity and kinetics). This method significantly improves single spike inference with CASCADE from calcium imaging data. Furthermore, we suggest the critical conditions required for developing robust auto-calibration approaches in the future.